A method for generating a spatio-temporally controllable remote sensing image simulation based on geographical meta-information constraints

Abstract

The application relates to the technical field of remote sensing images, and discloses a spatio-temporal controllable remote sensing image simulation generation method based on geographic meta-information constraints, which comprises the following steps: acquiring terrain elevation values, pixel spatial coordinates and remote sensing images containing radiation brightness values, spectral band characteristics and collection dates, performing feature mapping to obtain a geographic information vector; performing semantic coding according to the geographic information vector to obtain a regional feature vector and a trajectory feature vector; performing feature space alignment calculation according to the regional feature vector and the trajectory feature vector to obtain a fusion vector group; performing feature strengthening according to the fusion vector group to obtain a weighted feature matrix; performing pixel reconstruction according to the weighted feature matrix to obtain a remote sensing image sequence; performing iterative optimization according to the remote sensing image sequence to obtain an optimized sequence; and performing iterative correction according to the optimized sequence to obtain a final remote sensing image. The method can improve the spatio-temporal controllability and disaster scene adaptability of remote sensing image simulation generation.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image technology, and in particular to a spatiotemporally controllable remote sensing image simulation generation method based on geographic metadata constraints. Background Technology

[0002] Currently, remote sensing imagery needs to consider both spatial location and temporal changes in dynamic analysis scenarios such as agricultural monitoring, urban planning, and disaster early warning. However, the multidimensional nature of geographic environmental data and textual environmental descriptions are difficult to coordinate, resulting in insufficient accuracy in the spatiotemporal representation of the generated images. For example, when simulating flood disasters, geographic features and the severity of the disaster described in the text often fail to correspond accurately, affecting the practicality of the imagery. Therefore, it is necessary to introduce image enhancement techniques to improve the fusion capability of geographic information and textual descriptions.

[0003] In one existing technology, multispectral remote sensing images of a target area are acquired, and radiance values ​​and multi-band reflectance data are extracted. Simultaneously, topographic elevation data and pixel geographic coordinates are also acquired. This data is input into a convolutional neural network for feature extraction, yielding a high-dimensional feature vector representing the land cover type. Textual descriptions for the corresponding time period are acquired, and a word embedding model is used to convert the text into feature vectors. The image acquisition date is converted into a time index and concatenated with the aforementioned two feature vectors to obtain a comprehensive feature vector, which is then input into a generator of a generative adversarial network to generate a remote sensing image. However, this existing technology only performs simple data concatenation and static image generation, making it difficult to achieve deep collaboration between geographic information and textual descriptions.

[0004] Therefore, existing technology results in images that are not accurate enough in terms of spatiotemporal representation. Summary of the Invention

[0005] This invention provides a spatiotemporally controllable remote sensing image simulation generation method based on geographic metadata constraints, in order to improve the spatiotemporal controllability and disaster scenario adaptability of remote sensing image simulation generation.

[0006] In a first aspect, to address the aforementioned technical problems, this invention provides a spatiotemporally controllable remote sensing image simulation generation method based on geographic metadata constraints, comprising:

[0007] Acquire topographic elevation values, pixel spatial coordinates, and remote sensing images containing radiance values, spectral band characteristics, and acquisition dates;

[0008] Based on the remote sensing image, the terrain elevation value, and the pixel spatial coordinates, feature encoding is performed to obtain an abstract feature map, and feature mapping is performed based on the abstract feature map to obtain a geographic information vector.

[0009] Based on the geographic information vector, disaster association retrieval is performed through a preset disaster text database to obtain a disaster text report. Semantic encoding is then performed on the disaster text report to obtain regional feature vectors and trajectory feature vectors.

[0010] Based on the region feature vector and the trajectory feature vector, dimensional adjustment is performed to obtain the projected spectral vector. Then, based on the projected spectral vector, the terrain elevation value, and the trajectory feature vector, feature space alignment calculation is performed to obtain the fused vector group.

[0011] Based on the fusion vector group, the association weight is calculated to obtain the association weight distribution. Based on the association weight distribution, feature enhancement is performed based on the weight allocation judgment to obtain the weighted feature matrix.

[0012] Based on the weighted feature matrix, an image is generated using a pre-constructed generative adversarial network model to obtain an image feature map. Then, based on the image feature map, pixel reconstruction is performed based on spatial resolution consistency to obtain a remote sensing image sequence.

[0013] Based on the remote sensing image sequence, interruption point detection and statistics are performed to obtain the number of interruption points. Based on the number of interruption points, the remote sensing image sequence is iteratively optimized to obtain an optimized sequence.

[0014] Based on the optimized sequence, geometric registration verification is performed to generate a registration deviation matrix. Then, based on the registration deviation matrix and error judgment, iterative correction is performed to obtain the final remote sensing image.

[0015] Secondly, the present invention provides a spatiotemporally controllable remote sensing image simulation generation system based on geographic metadata constraints, comprising:

[0016] The data acquisition module is used to acquire topographic elevation values, pixel spatial coordinates, and remote sensing images containing radiance values, spectral band characteristics, and acquisition dates.

[0017] The mapping module is used to perform feature encoding based on the remote sensing image, the terrain elevation value and the pixel spatial coordinates to obtain an abstract feature map, and to perform feature mapping based on the abstract feature map to obtain a geographic information vector.

[0018] The encoding module is used to perform disaster association retrieval through a preset disaster text database based on the geographic information vector to obtain a disaster text report, and to perform semantic encoding based on the disaster text report to obtain regional feature vectors and trajectory feature vectors;

[0019] The alignment module is used to perform dimensional adjustment based on the region feature vector and the trajectory feature vector to obtain the projected spectral vector, and to perform feature space alignment calculation based on the projected spectral vector, the terrain elevation value and the trajectory feature vector to obtain the fused vector group;

[0020] The enhancement module is used to calculate the association weights based on the fusion vector group, obtain the association weight distribution, and perform feature enhancement based on the association weight distribution and weight allocation judgment to obtain a weighted feature matrix.

[0021] The reconstruction module is used to generate images based on the weighted feature matrix using a pre-built generative adversarial network model to obtain image feature maps, and to reconstruct pixels based on the image feature maps and spatial resolution consistency to obtain a remote sensing image sequence.

[0022] The optimization module is used to perform interruption point detection and statistics based on the remote sensing image sequence, obtain the number of interruption points, and perform iterative optimization on the remote sensing image sequence based on the number of interruption points to obtain an optimized sequence.

[0023] The output module is used to perform geometric registration verification based on the optimized sequence, generate a registration deviation matrix, and perform iterative correction based on the error judgment according to the registration deviation matrix to obtain the final remote sensing image.

[0024] Compared with the prior art, the present invention has the following beneficial effects:

[0025] (1) This invention extracts the radiance value, spectral band features and acquisition date of remote sensing images, combines the terrain elevation and pixel spatial coordinates, and uses a convolutional neural network to encode and map features to generate a geographic information vector with a time series index. This achieves deep fusion of heterogeneous data and unification of spatiotemporal dimensions, providing an accurate and continuous feature basis for subsequent dynamic analysis and disaster scenario simulation.

[0026] (2) This invention uses the BERT model to semantically encode disaster text, obtain regional feature vectors and trajectory feature vectors, and aligns feature dimensions through linear projection layers, calculates alignment points and residual vectors, and then introduces scaling dot product attention mechanism to strengthen highly correlated texture features, thereby improving the spatiotemporal matching degree between remote sensing images and text descriptions, and enhancing the adaptability of images to disaster scenes and the ability to express information.

[0027] (3) This invention generates image feature maps by generating adversarial networks and performs pixel reconstruction based on spatial resolution consistency constraints to obtain a preliminary remote sensing image sequence; further, iterative optimization is performed by interruption point detection and spectral fidelity loss function weight adjustment, and then registration deviation verification and iterative correction are performed by combining geometric correction model, and finally outputs spatiotemporally continuous and geographically accurate remote sensing images, which improves the authenticity and spatiotemporal controllability of image generation. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the spatiotemporally controllable remote sensing image simulation generation method based on geographic metadata constraints provided in the first embodiment of the present invention;

[0029] Figure 2 This is a schematic diagram of the spatiotemporally controllable remote sensing image simulation generation system based on geographic metadata constraints provided in the second embodiment of the present invention. Detailed Implementation

[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] Reference Figure 1 The first embodiment of the present invention provides a spatiotemporally controllable remote sensing image simulation generation method based on geographic metadata constraints, comprising the following steps:

[0032] S11, acquire topographic elevation values, pixel spatial coordinates, and remote sensing images containing radiance values, spectral band features, and acquisition date;

[0033] S12, based on the remote sensing image, the terrain elevation value and the pixel spatial coordinates, feature encoding is performed to obtain an abstract feature map, and feature mapping is performed based on the abstract feature map to obtain a geographic information vector;

[0034] S13, Based on the geographic information vector, perform disaster association retrieval through a preset disaster text database to obtain a disaster text report, and perform semantic encoding based on the disaster text report to obtain a regional feature vector and a trajectory feature vector;

[0035] S14. Based on the region feature vector and the trajectory feature vector, perform dimensional adjustment to obtain the projection spectral vector, and perform feature space alignment calculation based on the projection spectral vector, the terrain elevation value and the trajectory feature vector to obtain the fused vector group.

[0036] S15, based on the fusion vector group, perform association weight calculation to obtain association weight distribution, and based on the association weight distribution, perform feature enhancement based on weight allocation judgment to obtain a weighted feature matrix;

[0037] S16. Based on the weighted feature matrix, an image is generated using a pre-constructed generative adversarial network model to obtain an image feature map. Based on the image feature map, pixel reconstruction is performed based on spatial resolution consistency to obtain a remote sensing image sequence.

[0038] S17. Based on the remote sensing image sequence, perform interruption point detection and statistics to obtain the number of interruption points, and based on the number of interruption points, perform iterative optimization on the remote sensing image sequence to obtain an optimized sequence.

[0039] S18. Based on the optimized sequence, perform geometric registration verification to generate a registration deviation matrix. Based on the registration deviation matrix, perform iterative correction based on error judgment to obtain the final remote sensing image.

[0040] In step S11, the topographic elevation values, pixel spatial coordinates, and remote sensing images containing radiance values, spectral band characteristics, and acquisition dates are obtained.

[0041] Specifically, multispectral remote sensing image data covering the target area is read from publicly available or self-built remote sensing image databases. This image data contains the radiance values ​​of each pixel in multiple spectral bands (e.g., blue, green, red, and near-infrared bands), with the radiance values ​​recorded as floating-point numbers representing the electromagnetic radiation intensity received by each pixel. Simultaneously, the acquisition date is extracted from the image metadata, recorded in a standard date format for subsequent time-series indexing. Topographic elevation data consistent with the geographical extent of the remote sensing image is obtained from a digital elevation model database. This topographic elevation data, recorded as floating-point numbers, represents the altitude of each pixel location and originates from space shuttle radar topographic mapping missions or similar mapping results. The spatial coordinates of the pixels corresponding to the remote sensing image are obtained from a geographic information system. The spatial coordinates represent the geographical location of each pixel using a geographic coordinate system, with the coordinate values ​​stored as double-precision floating-point numbers. The obtained radiance values, spectral band features, acquisition dates, topographic elevation values, and pixel spatial coordinates are organized according to a one-to-one correspondence between pixels to form a structured input dataset. This step provides a complete raw data foundation for subsequent feature encoding and fusion, ensuring that spectral information, spatial location, terrain features, and temporal dimensions can be processed collaboratively within the same framework, thereby supporting the spatiotemporal accuracy and geographic realism of the generated images.

[0042] In step S12, feature encoding is performed based on the remote sensing image, the terrain elevation values, and the pixel spatial coordinates to obtain an abstract feature map. Then, feature mapping is performed based on the abstract feature map to obtain a geographic information vector, including:

[0043] Based on the radiance value and spectral band features in the remote sensing image, as well as the terrain elevation value and the pixel spatial coordinates, feature stitching is performed to generate a hybrid feature tensor.

[0044] Based on the hybrid feature tensor, feature encoding is performed using a pre-built convolutional neural network model to obtain an abstract feature map;

[0045] Based on the abstract feature map, feature mapping is performed using pre-constructed map mapping rules to obtain high-dimensional feature vectors;

[0046] Based on the acquisition date in the remote sensing image, a time series index is obtained by numerical conversion using a pre-constructed time axis encoding rule;

[0047] Based on the high-dimensional feature vector and the time series index, feature concatenation is performed to generate a geographic information vector containing the time series index, land cover type, and spectral feature vector.

[0048] Specifically, firstly, the radiance value and spectral band features of each pixel are extracted from the remote sensing imagery acquired in step S11; topographic elevation values ​​are extracted from the topographic elevation data; and pixel spatial coordinates are extracted from the geographic information system. These four types of data are then concatenated along the feature dimension in a pixel-to-pixel correspondence order to generate a multidimensional hybrid feature tensor. Each spatial location in this tensor corresponds to a pixel, and the feature vector at each pixel location is composed of the radiance value, multiple spectral band reflectance, topographic elevation, and spatial coordinates arranged sequentially. This operation does not change the individual feature values, but only expands the feature dimension.

[0049] The generated hybrid feature tensor is input into a pre-built convolutional neural network model for feature encoding. This convolutional neural network model employs a classic structure containing multiple convolutional layers, pooling layers, and activation function layers. For example, the network structure is designed as follows: the first layer uses a convolutional layer with a kernel size of 7×7 and a stride of 2, producing 64 output channels, followed by batch normalization and rectified linear activation; the second layer uses a convolutional layer with a kernel size of 3×3 and a stride of 2, producing 128 output channels, followed by batch normalization and rectified linear activation; the third to fifth layers all use convolutional layers with a kernel size of 3×3 and a stride of 1, producing 256, 512, and 512 output channels respectively, followed by batch normalization and rectified linear activation; after each convolutional layer, a max pooling layer is used for spatial downsampling, with a pooling window size of 2×2 and a stride of 2. Finally, a global average pooling layer compresses the spatial dimension into a one-dimensional feature vector. The model construction process involves collecting a large number of images covering different land cover types and their corresponding land cover type labels from publicly available remote sensing image datasets. Data preprocessing steps include resampling all images to a uniform spatial resolution, converting radiance values ​​to reflectance values ​​and normalizing them. The normalization method involves calculating the mean and standard deviation for each band, subtracting the mean from each pixel value, and then dividing by the standard deviation. Data augmentation techniques such as random horizontal flipping, vertical flipping, and rotation are used to expand the training samples. Using a hybrid feature tensor as input and land cover type classification as the objective, the network parameters are trained using a cross-entropy loss function and a stochastic gradient descent optimization algorithm until the loss converges. After training, the model can map the input multidimensional hybrid feature tensor into an abstract feature map. This map is organized in the form of a two-dimensional matrix, with each spatial location corresponding to a feature vector. This vector encodes the dominant land cover type and texture structure information at that location.

[0050] An abstract feature map is input into a pre-constructed map mapping rule for feature mapping, resulting in a high-dimensional feature vector. The map mapping rule is implemented by a fully connected layer. The input dimension of this fully connected layer is equal to the dimension of the feature vector at each location in the abstract feature map, and the output dimension is a preset high-dimensional vector dimension (e.g., 128 dimensions). The weight matrix of the fully connected layer is obtained through training. While training the convolutional neural network, the fully connected layer is appended after the convolutional neural network, and joint optimization is performed using the same training data and classification task, ensuring that the high-dimensional feature vector output by the fully connected layer can represent land cover type and texture information. In practical use, the abstract feature map is flattened into a one-dimensional vector and input to the fully connected layer. Through matrix multiplication (i.e., multiplying the weight matrix with the input vector and adding a bias) and nonlinear transformation by an activation function (such as a linear rectified function), a fixed-length high-dimensional feature vector is obtained.

[0051] The acquisition date is extracted from the remote sensing image metadata obtained in step S11. This date is then input into a pre-constructed timeline encoding rule for numerical conversion to obtain a time series index. The timeline encoding rule is constructed by statistically analyzing the acquisition dates of all images in the historical remote sensing image dataset, determining the earliest and latest dates, dividing the entire time range into equally spaced time points, and assigning an integer index to each time point. For any given date, the number of days from the earliest date is calculated, divided by the total number of days, multiplied by a preset maximum index value (e.g., 1000), and then rounded to the nearest integer to obtain a unique time series index. This index value reflects the relative position of the image on the timeline, facilitating subsequent time series processing.

[0052] The obtained high-dimensional feature vector is concatenated with the time-series index. Specifically, the high-dimensional feature vector is expanded into a vector by adding a scalar index value, and then concatenated end-to-end to generate a geographic information vector. This geographic information vector contains spectral feature vectors from the high-dimensional feature vector, land cover type information, and an additional time-series index. This achieves the fusion of heterogeneous features in a unified vector space, providing a geographic feature foundation with clear time identification for subsequent analysis of disaster-related texts and spatiotemporal dynamics.

[0053] In step S13, based on the geographic information vector, a disaster association retrieval is performed through a preset disaster text database to obtain a disaster text report. Semantic encoding is then performed on the disaster text report to obtain regional feature vectors and trajectory feature vectors, including:

[0054] Based on the time series index and land cover type in the geographic information vector, a disaster text report is obtained by performing a correlation retrieval through a preset disaster text database;

[0055] Identify the geographical location nouns and state change verbs in the disaster text report, and extract the descriptive sentences based on the geographical location nouns and state change verbs through dependency parsing.

[0056] Based on the description, semantic encoding is performed using a pre-built BERT model to obtain region feature vectors and trajectory feature vectors.

[0057] Specifically, disaster text reports associated with the geographic information vector are retrieved from a pre-defined disaster text database. The disaster text database is constructed by collecting historical disaster event records covering the target area, including unstructured text data such as government disaster reports, news reports, and case descriptions in academic literature, stored in plain text format. Each record includes keywords for the event's time and location, as well as a disaster type tag. An inverted index is built into the text in the database, arranged chronologically, using the time field and disaster type field as search keys. During associated searches, the time series index and land cover type code are extracted from the geographic information vector. A search time window is determined based on the time series index, for example, extending 30 days forward and backward from the date corresponding to the index. All text records whose occurrence time falls within this window are selected from the disaster text database. Simultaneously, based on the disaster type corresponding to the land cover type code, texts containing keywords of that disaster type are further matched from the filtered results, ultimately yielding disaster text reports spatiotemporally associated with the geographic information vector.

[0058] After obtaining disaster text reports, natural language processing (NLP) is performed on the report text to extract key descriptive statements. First, the text is segmented and tagged with parts of speech to identify geographical nouns and verbs indicating state changes. Geographical nouns include administrative division names, natural feature names, and directional descriptions; state change verbs include verbs describing the disaster process and its impact. The identification process uses a pre-built named entity recognition model, which is based on a bidirectional long short-term memory network and a conditional random field (CRF). During model construction, a labeled disaster corpus is first established. Collected historical disaster event texts are cleaned, removing HTML tags, special symbols, and meaningless characters, followed by word segmentation and part-of-speech tagging. Then, manual tagging is performed, classifying each word as a geographical noun, a state change verb, or another category. Geographical nouns are further subdivided into administrative divisions, natural features, and directional descriptions; state change verbs are subdivided into disaster-related verbs such as flooding, fire, and collapse. The model uses a bidirectional long short-term memory network layer to capture the contextual features of the text sequence, and a CRF layer to model the dependencies between the tag sequences, outputting the optimal tag for each word. During training, the word vector sequence of the labeled text is input into a bidirectional long short-term memory network to obtain the feature representation of each word. Then, it is input into a conditional random field to calculate the score of the label sequence. Negative log-likelihood is used as the loss function. The network parameters are optimized through backpropagation and stochastic gradient descent. The training is iterative until the loss converges.

[0059] After training, the model can perform word-by-word annotation on any input text, outputting the entity category and location corresponding to each word, thereby extracting all geographic location nouns and state change verbs and their positions in the sentence. Then, using these words as anchors, dependency parsing is used to construct the sentence's grammatical dependency tree. Based on dependency relations, words that form subject-verb, verb-object, or modifier relationships with geographic location nouns and state change verbs are identified and combined to form complete descriptive sentences. For each disaster text report, multiple descriptive sentences can be extracted, corresponding to descriptions of the disaster area's extent and descriptions of land cover changes.

[0060] The extracted descriptive statements are input into a pre-built BERT model for semantic encoding, yielding regional and trajectory feature vectors. The BERT model is constructed using a pre-trained architecture with a twelve-layer Transformer encoder. Model parameters are pre-trained on a general corpus and fine-tuned on disaster-related corpora. During fine-tuning, a large amount of disaster text and its corresponding labels (disaster type, area range, change type, etc.) are collected. The descriptive statements are used as input, and a multi-task learning approach is employed to simultaneously predict regional and trajectory relevance. The cross-entropy loss function is optimized to enable the model to capture spatial and dynamic information from disaster texts. In data preprocessing, the disaster text is cleaned to remove noise characters. The BERT model's built-in word segmenter is used for word segmentation, with classification tags added to the beginning and delimiters to the end of the segmented sequences. The sequences are then padded to a uniform length and attention masks are generated. In practice, the descriptive statements are input into the BERT model. The average pooling of all word vectors in the last layer is used as the overall semantic vector. Simultaneously, specific positions are mapped through linear layers to obtain regional and trajectory feature vectors, respectively. Regional feature vectors encode spatial information such as the geographical location and extent of disasters, while trajectory feature vectors encode dynamic information such as the type, intensity, and temporal sequence of changes in the state of ground features.

[0061] In step S14, dimensional adjustment is performed based on the region feature vector and the trajectory feature vector to obtain a projected spectral vector. Then, feature space alignment calculation is performed based on the projected spectral vector, the terrain elevation value, and the trajectory feature vector to obtain a fused vector set, including:

[0062] When the dimensions of the spectral feature vector and the regional feature vector of the geographic information vector are inconsistent, the dimensional transformation of the spectral feature vector is performed through a pre-constructed linear projection layer to obtain the projected spectral vector;

[0063] When the spectral feature vector of the geographic information vector and the dimension of the regional feature vector are the same, the spectral feature vector is directly used as the projected spectral vector.

[0064] Based on the terrain elevation values, a terrain vector is obtained by performing dimensional mapping through a pre-constructed terrain mapping layer, and a trajectory vector is obtained by performing dimensional mapping through the linear projection layer based on the trajectory feature vector.

[0065] The alignment point is obtained by calculating the average vector of the projected spectral vector, the terrain vector, the regional feature vector, and the trajectory vector;

[0066] The differences between the projected spectral vector, the terrain vector, the regional feature vector, and the trajectory vector and the alignment point are calculated respectively, and the vectors are concatenated according to the calculation results to obtain a fused vector group containing texture features and text features.

[0067] Specifically, the spectral feature vector from the geographic information vector and the regional feature vector obtained from step S13 are acquired, and their feature dimensions are compared. The geographic information vector is generated in step S12, and its spectral feature vector is a part of a high-dimensional feature vector. The dimension of this spectral feature vector is determined by the output dimension of the fully connected layer in step S12 (e.g., 128 dimensions). The regional feature vector is obtained by mapping the BERT model output in step S13 through a linear layer, and its dimension is consistent with the dimension set by the BERT model output layer. If the dimensions of the two vectors are different, the spectral feature vector needs to be transformed in dimension, which is done using a pre-constructed linear projection layer.

[0068] The linear projection layer is a fully connected neural network layer. Its input dimension is equal to the dimension of the spectral feature vector, and its output dimension is equal to the dimension of the regional feature vector. The construction process of the linear projection layer is as follows: a large number of samples (e.g., 100,000) are randomly selected from a publicly available remote sensing image dataset. Each sample contains a spectral feature vector and its corresponding regional feature vector. The spectral feature vector is extracted from the image using the method described in step S12, and the regional feature vector is encoded from the corresponding disaster text using the method described in step S13. The spectral feature vectors of all samples are used to form an input matrix, with each row being a spectral feature vector. The corresponding regional feature vectors are used to form a target matrix, with each row being a regional feature vector. The weight matrix of the linear projection layer is initialized to a random value with a mean of zero and a standard deviation equal to the reciprocal of the square root of the input dimension. The bias vector is initialized to zero. During training, one hundred samples are randomly selected from the input matrix to form a mini-batch. These one hundred spectral feature vectors are multiplied by the weight matrix to obtain one hundred output vectors. These output vectors are then added to the bias vector to obtain the prediction vector for that batch. The mean squared error between the predicted and target vectors is calculated as the loss value. This mean squared error is equal to the sum of the squares of the differences between the predicted and target values ​​in each dimension, divided by the vector dimension and then by the batch size. The gradients of the weight matrix and bias vector are calculated based on the loss value. Each element of the weight matrix is ​​subtracted from the product of the learning rate and the corresponding gradient, and the same applies to each element of the bias vector. The learning rate is initially set to 0.001. The above sampling, forward computation, loss calculation, and gradient update steps are repeated. Each iteration after processing all samples constitutes one training epoch, for a total of fifty epochs. After training, the final weight matrix and bias vector are obtained. In use, the spectral feature vector is multiplied by the weight matrix and the bias is added to obtain the projected spectral vector. If both dimensions are the same, the spectral feature vector is directly used as the projected spectral vector without transformation.

[0069] The process involves processing topographic elevation values, regional feature vectors, and trajectory feature vectors. The topographic elevation values ​​are the elevation scalar values ​​of each pixel obtained in step S11, which are input to a pre-constructed topographic mapping layer for dimension mapping. The topographic mapping layer is a fully connected layer with an input dimension of 1 and an output dimension set to the same dimension as the projected spectral vector. This mapping layer is constructed by collecting a large number of samples (e.g., 100,000), each containing a topographic elevation value and a corresponding spectral feature vector extracted from remote sensing images of the same location. The topographic elevation values ​​form the input matrix, and the spectral feature vectors form the target matrix. The weight matrix of the topographic mapping layer is initialized with random values, and the bias vector is initialized to zero. The training process is the same as that of the linear projection layer, using stochastic gradient descent to optimize the mean squared error loss. The final weights and biases are obtained after fifty training iterations. In use, the topographic elevation value of each pixel is input to this layer, multiplied by the weight matrix and then added with the bias, followed by activation by a linear rectified function to obtain a topographic vector with the same dimension as the projected spectral vector. Simultaneously, dimensional mapping is performed on the trajectory feature vector and region feature vector. These vectors are output by the BERT model in step S13, and their dimensions may differ from the projected spectral vector. Therefore, the same linear projection layers as described above are used for dimensional transformation. The input dimensions of these two linear projection layers are the original dimensions of the trajectory feature vector and region feature vector, respectively, and the output dimensions are set to the dimensions of the projected spectral vector. During training, the trajectory feature vector and region feature vector are used as inputs, and the projected spectral vector is used as the output target. After fifty rounds of stochastic gradient descent training, their respective weight matrices and biases are obtained. When used, the trajectory feature vector and region feature vector are input to the corresponding layers to obtain trajectory vectors and region vectors with the same dimensions as the projected spectral vector. It should be noted that the trajectory vector and region vector are global features, with only one vector corresponding to each image region; while the projected spectral vector and terrain vector are pixel-wise features, with one vector corresponding to each pixel position. To perform pixel-wise alignment calculations, the trajectory vector and region vector are first copied along the spatial dimension to make their spatial dimensions consistent with the projected spectral vector, meaning that each pixel position has the same trajectory vector and region vector.

[0070] After obtaining four vectors of the same dimension, for each pixel location, the projected spectral vector, terrain vector, copied trajectory vector, and region vector at that location are grouped into a set of four vectors. The average of these four vectors is calculated as the alignment point. Specifically, the four vectors are added element-wise to obtain the sum vector in each dimension, then divided by 4 to obtain the average value in each dimension, forming a new vector, which is the alignment point. Subsequently, the differences between the projected spectral vector, terrain vector, copied region vector, and trajectory vector and the alignment point are calculated. That is, for each dimension, the element value of the projection spectral vector is subtracted from the element value of the corresponding dimension of the alignment point to obtain the projection spectral residual vector; similarly, the terrain residual vector, region residual vector, and trajectory residual vector are obtained. These four residual vectors are concatenated end-to-end in the feature dimension to form a long vector with a dimension four times that of the original dimension, which serves as the fusion feature for that pixel location. After performing the above operation on all pixel locations, a fusion vector group containing texture features, region features, and text features is obtained. This fusion vector group is a three-dimensional tensor with the same spatial dimension as the original image, and each spatial location corresponds to a fusion feature vector.

[0071] In step S15, based on the fusion vector group, association weights are calculated to obtain an association weight distribution. Then, based on the association weight distribution and weight allocation judgment, feature enhancement is performed to obtain a weighted feature matrix, including:

[0072] Based on the text features in the fused vector group, an attention query vector is obtained by linear transformation through a pre-constructed query projection layer.

[0073] Based on the texture features in the fused vector group, linear transformations are performed through a pre-constructed key projection layer and value projection layer to obtain the key matrix and value matrix, respectively.

[0074] Based on the attention query vector and the key matrix, the weights are calculated using a scaled dot product attention mechanism to obtain the associated weight distribution;

[0075] Based on the associated weight distribution, the texture features and weight values ​​corresponding to the dimensions whose weight values ​​exceed a preset weight threshold are marked to obtain highly associated feature pairs;

[0076] Based on the highly correlated features, an amplification factor is calculated using a preset exponential mapping function to obtain a feature enhancement factor. The feature enhancement factor and the value matrix are then multiplied by weight to obtain a weighted feature matrix.

[0077] Specifically, text features and texture features are extracted from the fusion vector group generated in step S14. The fusion vector group is composed of four residual vectors concatenated. The text features are derived from the semantic information carried by the trajectory residual vector and the region residual vector, while the texture features are derived from the surface structure information carried by the projection spectral residual vector and the terrain residual vector. The text features are input to a pre-constructed query projection layer for linear transformation to obtain the attention query vector. The query projection layer is a fully connected neural network layer with an input dimension equal to the dimension of the text features and an output dimension set to a preset attention head dimension (e.g., 64-dimensional). The construction process of this query projection layer is the same as that of the linear projection layer in step S14, i.e., randomly sampling samples (e.g., 100,000) from the training dataset, each sample containing a text feature vector, using the text feature vector as input, and a randomly initialized target vector as the supervision signal. Stochastic gradient descent is used to optimize the mean squared error loss. One hundred samples are randomly sampled each time to form a mini-batch, and the mean squared error between the predicted vector and the target vector is calculated. The gradient is calculated based on the error, and the weight matrix and bias are updated. After fifty training epochs, the final weight matrix and bias are obtained. When using it, the text feature vector is multiplied by the weight matrix and a bias is added to obtain the attention query vector.

[0078] The texture features from the fused vector group are input into pre-constructed key projection and value projection layers for linear transformation, yielding key and value matrices, respectively. Both key and value projection layers are fully connected layers, with input dimensions equal to the texture feature dimensions and output dimensions identical to the query vector. The construction process for these two projection layers is the same as that for the linear projection layer in step S14, using texture feature vectors as input and randomly initialized target vectors as supervision signals. They are trained for fifty epochs using the same stochastic gradient descent method to obtain their respective weight matrices and biases. In use, the texture feature vectors are multiplied by the weight matrices of the key and value projection layers, and a bias is added to obtain key and value vectors. All key vectors are arranged sequentially to form the key matrix, and all value vectors are arranged sequentially to form the value matrix.

[0079] After obtaining the attention query vector, key matrix, and value matrix, the association weight distribution is calculated using a scaled dot product attention mechanism. Specifically, the attention query vector and the key matrix are multiplied by the corresponding element of each key vector in the key matrix, and the results are summed to obtain a similarity score vector between the query vector and each key vector. Then, each element of this similarity score vector is divided by the square root of the query vector's dimension (e.g., by 8) to scale it. The scaled scores are then input into a normalized exponential function, which calculates an exponent with base e for each score and divides it by the sum of all exponents, transforming all scores into a probability distribution with a sum of 1, which is the association weight distribution.

[0080] Based on the association weight distribution, texture features and weight values ​​corresponding to dimensions with weight values ​​exceeding a preset weight threshold are marked to obtain highly correlated feature pairs. The weight threshold is determined by collecting one thousand labeled fusion vector groups and their corresponding ideal association weights from historical disaster monitoring datasets. These ideal association weights originate from verified disaster event samples in the historical disaster monitoring datasets, whose disaster area boundaries have been accurately determined through high-precision remote sensing image interpretation or field survey data. Steps S13 to S14 are performed on each sample group to obtain fusion vector groups. Then, an attention mechanism is executed to calculate the association weight distribution, and the weight values ​​of texture feature dimensions corresponding to verified disaster areas in the distribution are used as the ideal association weights for that group. After calculating the association weight distribution for each data group, the weight values ​​of all positive samples are counted, and these weight values ​​are sorted from smallest to largest. The value at the 900th position is taken as the threshold (e.g., 0.6). This threshold is used to filter the current association weight distribution. For each dimension, if the weight value of that dimension is greater than the threshold, the texture feature vector and its weight value corresponding to that dimension are extracted to form a set of highly correlated feature pairs.

[0081] For each highly correlated feature pair, a feature enhancement factor is calculated based on its weight value using a predefined exponential mapping function. The exponential mapping function is constructed as follows: the weight value is the independent variable, the magnification factor is the dependent variable, and the function form is an exponential function with the natural constant e as the base. The exponent is the difference between the weight value and the baseline value, divided by a temperature coefficient. The baseline value is set as a threshold (e.g., 0.6), and the temperature coefficient is set as an empirical value (e.g., 0.2). This temperature coefficient can be obtained by performing a grid search on the validation set. Specifically, different temperature coefficients are tried in a step size of 0.05 within the range of 0.1 to 0.5, and the value that optimizes the generated image quality index, such as the peak signal-to-noise ratio, is selected as the final temperature coefficient. In the specific calculation, the weight value is subtracted from the baseline value, divided by the temperature coefficient, and then the exponent with base e is calculated to obtain the enhancement factor. The calculated enhancement factor is then multiplied element-wise by the value vector corresponding to that feature in the value matrix, i.e., the enhancement factor is multiplied by each element in the value vector to obtain the enhanced feature vector. After performing this operation on all highly correlated feature pairs, the feature vectors at the corresponding positions in the value matrix are enhanced, while other positions remain unchanged, forming an enhanced feature matrix, which is the weighted feature matrix.

[0082] In step S16, based on the weighted feature matrix, an image is generated using a pre-constructed generative adversarial network model to obtain an image feature map. Then, based on the image feature map, pixel reconstruction is performed according to spatial resolution consistency to obtain a remote sensing image sequence, including:

[0083] Based on the time series index, the weighted feature matrix is ​​sliced ​​in time series to obtain a time phase slice group. Based on the time phase slice group, spatial features are aggregated by global average pooling to obtain a time phase feature vector group.

[0084] The difference and cosine similarity between adjacent time-phase feature vectors in the time-phase feature vector group are calculated, and image features are generated through a pre-constructed generative adversarial network model based on the calculation results to obtain an image feature map.

[0085] When the image feature map satisfies the preset spatial resolution consistency constraint, the spatial pixel dimension is transformed through the pre-constructed pixel reconstruction mapping layer to obtain the pixel reconstruction matrix;

[0086] When the image feature map does not satisfy the spatial resolution consistency constraint, the parameters of the generative adversarial network model are adjusted, and the image feature map is regenerated until the image feature map satisfies the spatial resolution consistency constraint, thus obtaining the pixel reconstruction matrix;

[0087] Based on the pixel reconstruction matrix and the image feature map, pixel transformation is performed through matrix multiplication to obtain a remote sensing image sequence.

[0088] Specifically, the time-series index set generated in step S12 is obtained. This set contains index values ​​for multiple historical moments, with each index value uniquely corresponding to a historical moment. For each time-series index, steps S13 to S15 are executed to obtain the weighted feature matrix corresponding to that moment. The weighted feature matrices of all moments are stacked in chronological order to form a three-dimensional tensor, with dimensions of time-series length, spatial height, and spatial width, and each spatial location containing a feature vector. This yields a weighted feature tensor with a time dimension. The time-series index is extracted from this three-dimensional tensor. It should be noted that the time-series index originates from the value converted from the original remote sensing image acquisition date in step S12. Its function is to serve as a time anchor point during the generation process, ensuring that the generated image sequence corresponds to the specified historical moment, thereby achieving controllability of the time dimension. The weighted feature matrix is ​​segmented in the time dimension according to the time-series index, with each time point corresponding to a two-dimensional spatial feature matrix. The feature matrices of all time points are arranged in chronological order to form a temporal slice group. For example, if a time series index contains twelve time points, then twelve spatial feature matrices are obtained, each matrix having a size equal to its height multiplied by its width multiplied by the number of feature channels.

[0089] Spatial feature aggregation is performed on each temporal slice in the temporal slice group using global average pooling. Specifically, for each temporal slice, the values ​​of all spatial locations in each feature channel are summed and divided by the total number of spatial locations to obtain the average value for that channel. After performing this operation on all feature channels, each temporal slice is compressed into a one-dimensional feature vector with a length equal to the number of feature channels. The feature vectors corresponding to all temporal slices are arranged in chronological order to form a temporal feature vector group. This vector group is a two-dimensional matrix, with the number of rows representing the number of time points and the number of columns representing the number of feature channels.

[0090] Calculate the difference and cosine similarity between adjacent time-phase feature vectors in the time-phase feature vector group. For each adjacent time point pair, subtract the feature vector of the previous time point from the feature vector of the later time point to obtain the difference vector. All difference vectors are arranged in chronological order to form a change trajectory data matrix. Simultaneously, calculate the cosine similarity between adjacent feature vectors by multiplying corresponding elements of two feature vectors, summing the results, and then dividing by the product of the magnitudes of the two vectors. This yields a similarity value between negative one and positive one. All similarity values ​​are arranged in chronological order to form a continuous data sequence.

[0091] The changing trajectory data matrix and continuous data sequence are used as inputs to a pre-constructed generative adversarial network (GAN) model for image feature generation. The GAN model consists of a generator and a discriminator. The generator employs a deconvolutional neural network (DCN) structure, with its input layer dimension equal to the length of a one-dimensional vector concatenated from the changing trajectory data matrix and continuous data sequence. The spatial size is progressively expanded through multiple deconvolutional layers, ultimately outputting a three-dimensional image feature map whose spatial size matches that of the target remote sensing image. The number of feature channels is set to a preset value. The discriminator employs a convolutional neural network (CNN) structure, taking either the image feature map output by the generator or a real image feature map as input, and outputting a scalar representing the probability of the input being genuine. The model construction process involves preprocessing the collected changing trajectory data matrix and continuous data sequence. For each sample's changing trajectory data matrix, the mean and standard deviation are calculated for each channel. Each element is standardized by subtracting the corresponding channel mean and dividing by the standard deviation. Similarly, the global mean and standard deviation are calculated for the continuous data sequence to ensure a uniform numerical range for all input data. Ten thousand samples containing real change trajectories and corresponding real images were collected from a historical remote sensing image database. The real change trajectory data was input into a generator to obtain generated image feature maps. The generated image feature maps and real image feature maps were then input into a discriminator, and an adversarial training method was used for optimization. The generator's loss function includes adversarial loss and feature matching loss. The adversarial loss aims to make the generated image feature maps fool the discriminator, while the feature matching loss minimizes the mean squared error between the generated and real image feature maps. The discriminator's loss function is the cross-entropy loss, which distinguishes between real image feature maps as true and generated image feature maps as false. The Adam optimization algorithm was used for parameter updates. In each training epoch, for each mini-batch (batch size set to 32), the gradient of the loss function with respect to the current model parameters was first calculated. Then, the first-moment and second-moment estimates of the gradient were calculated separately. The first-moment estimate was the weighted average of the current gradient and the first-moment estimate from the previous time step, with a weighting coefficient of 0.9. The second-moment estimate was the weighted average of the squared current gradient and the second-moment estimate from the previous time step, with a weighting coefficient of 0.999. Next, bias correction is applied to the first-order and second-order moment estimates. The correction method involves dividing the first-order moment estimate by one minus 0.9 raised to the power of the current iteration number, and dividing the second-order moment estimate by one minus 0.999 raised to the power of the current iteration number. Finally, the corrected first-order moment estimate is divided by the square root of the corrected second-order moment estimate, and then multiplied by the learning rate of 0.0002 to obtain the parameter update amount. This update amount is then subtracted from the current parameters to complete one parameter update. The generator and discriminator parameters are updated alternately, and the above process is repeated until two hundred training epochs are completed. After training, the generator can generate image feature maps that conform to the temporal evolution pattern based on the input change trajectory and continuous data. In use, the change trajectory data matrix and continuous data sequence obtained in the previous steps are concatenated and input into the generator, and the image feature maps are obtained through forward propagation.

[0092] Spatial resolution consistency is checked on the generated image feature maps. The process for determining the spatial resolution consistency constraint is as follows: First, one thousand high-resolution real remote sensing images are randomly selected from the training dataset. Each image's corresponding feature map is calculated using the same feature extraction network, resulting in a set of reference feature maps. Based on the spatial resolution of the target remote sensing image, the corresponding spatial cutoff frequency is calculated. This cutoff frequency is equal to half the reciprocal of the ground sampling distance. A two-dimensional discrete Fourier transform is performed on each reference feature map to obtain its spectrum. The total energy of components with frequencies higher than the cutoff frequency is calculated and divided by the total energy to obtain the high-frequency energy percentage of each reference feature map. These one thousand high-frequency energy percentages are sorted from smallest to largest, and the value at the 950th position is taken as the high-frequency energy threshold. When checking the current image feature map, its high-frequency energy percentage is also calculated. If this percentage is not lower than the above threshold, the spatial resolution consistency constraint is considered satisfied; otherwise, it is considered not satisfied. If the constraints are not met, the parameters of the generative adversarial network model are adjusted. Specifically, the current image feature map is input into the discriminator, and the generator parameters are updated using the gradient output by the discriminator. That is, the gradient of the loss with respect to the generator parameters is calculated through backpropagation, and the generator parameters are updated with a gradient descent step size of 0.00001 with a learning rate. Then, the updated generator is used to regenerate the image feature map, and the test is performed again. This process is repeated until the generated image feature map satisfies the constraints, and the generator parameters when the constraints are satisfied are recorded.

[0093] After satisfying the spatial resolution consistency constraint, the image feature map is input to a pre-constructed pixel reconstruction mapping layer for spatial pixel dimension transformation to obtain the pixel reconstruction matrix. The pixel reconstruction mapping layer is a fully connected layer whose input dimension equals the number of feature channels in the image feature map, and whose output dimension equals the number of bands in the target remote sensing image. The construction process of this mapping layer involves collecting 10,000 sets of image feature maps and corresponding real remote sensing image pixel values ​​from the training dataset. The image feature maps are generated by the generator, and the real pixel values ​​are directly extracted from the images. The spatial location of the image feature map is treated as an independent sample. The feature vector at each location is used as the input, and the pixel value vector at that location is used as the output target. The fully connected layer is trained using the mean squared error loss function and stochastic gradient descent optimization algorithm. One hundred samples are randomly selected each time, and the weight matrix and bias are obtained after fifty training iterations. In use, the feature vector of each spatial location in the image feature map is multiplied by the weight matrix and a bias is added to obtain the pixel value vector at that location. The pixel value vectors of all locations are arranged according to their spatial location to form a pixel reconstruction matrix. The dimension of this matrix is ​​the spatial height multiplied by the spatial width multiplied by the number of bands, which corresponds to a multi-band remote sensing image.

[0094] Pixel transformation is achieved by multiplying the pixel reconstruction matrix with the image feature map to obtain a remote sensing image sequence. Specifically, the image feature map is three-dimensional, and the pixel reconstruction matrix is ​​essentially a linear transformation that maps feature channels to band numbers, applied to each spatial location. Therefore, for each spatial location, the feature vector at that location is multiplied by the weight matrix of the pixel reconstruction matrix to obtain the band value vector at that location. After performing this operation on all spatial locations, a three-dimensional remote sensing image is obtained. If a temporal slice group contains multiple time points, the above transformation is performed on the image feature map at each time point separately to obtain remote sensing images at multiple time points, which are then arranged in chronological order to form a remote sensing image sequence.

[0095] In step S17, based on the remote sensing image sequence, interruption point detection and statistics are performed to obtain the number of interruption points. Based on the number of interruption points, the remote sensing image sequence is iteratively optimized to obtain an optimized sequence, including:

[0096] Based on the remote sensing image sequence and the preset reference surface reflectance, the difference between the pixel value and the reference value is calculated pixel by pixel to obtain the error matrix;

[0097] Based on the error matrix, breakpoints are detected using a pre-built smoothness detection model to obtain the number of breakpoints.

[0098] When the number of interruption points is not higher than the preset threshold for the number of interruption points, the remote sensing image sequence is directly used as the optimized sequence.

[0099] When the number of interruption points exceeds the interruption point threshold, the weights of the spectral fidelity loss function in the generative adversarial network model are adjusted, and the image feature map generation step is re-executed until the number of interruption points is no higher than the interruption point threshold, thus obtaining an optimized sequence.

[0100] Specifically, the radiance value of each pixel is obtained from the remote sensing image sequence generated in step S16, and real surface reflectance data with the same spatial range and temporal phase as the remote sensing image sequence is obtained from a preset reference surface reflectance database. The reference surface reflectance database is constructed by collecting historical high-precision remote sensing images covering the target area. These images have undergone radiometric calibration and atmospheric correction, and the reflectance value of each pixel is used as a reference standard, stored by geographic location and time index. For each pixel in the remote sensing image sequence, the radiance value of the generated pixel is subtracted from the reference reflectance value at the corresponding location to obtain the difference. The absolute value of the difference is the error of that pixel. After performing this operation on all pixels, a two-dimensional error matrix with the same spatial size as the remote sensing image sequence is obtained. The value of each element in the matrix represents the degree of deviation between the generated image and the real reflectance at that location.

[0101] The error matrix is ​​input into a pre-built smoothness detection model for interruption point detection. The smoothness detection model employs a one-dimensional time-series convolutional neural network structure. Its construction process involves collecting one thousand samples containing continuous temporal error matrices from historical disaster monitoring datasets. Each sample consists of error matrices from five consecutive temporal phases, with the size of each phase's error matrix being the height multiplied by its width. The error values ​​of each pixel are arranged chronologically to form the pixel's time series. During data preprocessing, the time series of each pixel is standardized by calculating its mean and standard deviation. The error value at each time step is subtracted from the mean and divided by the standard deviation, ensuring the input data has zero mean and unit variance. The gradient of change for each pixel's time series is calculated, representing the difference between adjacent temporal error values. This gradient is used as a label, marking time points where the absolute value of the gradient exceeds a preset gradient threshold (e.g., 0.1) as interruption points. Using the time series of all pixels in the sample as input and the corresponding interruption point labels as output, a one-dimensional convolutional neural network is trained. This network consists of three convolutional layers, each followed by batch normalization and a linear rectified activation function, and finally a fully connected layer and a sigmoid activation function. It outputs the probability of whether each time point is an interruption point. During training, a binary cross-entropy loss function is used, optimized using the Adam algorithm, with a learning rate of 0.001, a batch size of 64, and 100 training epochs. After training, the model can predict whether each time point is an interruption point based on the input time series error values. In use, the error values ​​of each pixel in the error matrix are arranged chronologically into a time series, input into the model, and the prediction of whether each pixel is interrupted at each time point is obtained. The interruption point predictions for all pixels are summed to obtain the total number of interruption points in the entire sequence.

[0102] The method for determining the threshold for the number of interruptions is as follows: One thousand sets of real remote sensing image sequences without disasters are selected from a reference surface reflectance database. Each set contains images from five consecutive time phases. The absolute value of the reflectance difference for each pixel between adjacent time phases is calculated for each set. The absolute values ​​of the differences for all sequences, all time phases, and all pixels are summed to obtain a set of difference values. Statistical analysis is performed on this set, and its cumulative distribution function is calculated. The 99th percentile is taken as the preset difference threshold (e.g., 0.05). This threshold indicates that under normal circumstances, 99% of pixel variations should be below this value. Then, five hundred sets of real remote sensing image sequences are randomly selected from the reference surface reflectance database. For each set, locations where the difference between adjacent time phases exceeds the preset threshold are marked as real interruptions. The number of interruptions in each set is counted. These five hundred sets of sequences are sorted from smallest to largest by the number of interruptions, and the value at the 95th position is taken as the threshold for the number of interruptions. This threshold indicates that under normal circumstances, 95% of high-quality image sequences will not exceed this value in terms of the number of interruptions.

[0103] The number of detected breakpoints is compared with a preset threshold. If the number of breakpoints is not higher than the threshold, the current remote sensing image sequence is directly output as the optimized sequence. If the number of breakpoints is higher than the threshold, the weights of the spectral fidelity loss function in the generative adversarial network (GAN) model need to be adjusted. It's important to note that this adjustment does not directly modify the trained model parameters during the model inference stage. Instead, it is achieved through an iterative fine-tuning mechanism. Using the current GAN model as the initial model, the weights of the spectral fidelity loss function are increased from the initial value of 0.5 to a new weight value. A small learning rate is used to fine-tune the generator parameters on a small amount of validation data for a limited number of rounds (e.g., five rounds), allowing the generator to maintain its original generation capabilities while enhancing its constraint on spectral fidelity. During fine-tuning, the discriminator parameters can remain fixed or be updated with the same small learning rate to avoid over-adjustment of the generator and compromise generation quality. The adjusted weights are used to recalculate the total loss function of the GAN, which is a weighted sum of the adversarial loss, feature matching loss, and spectral fidelity loss. Then, the image feature map generation step in step S16 is repeated. Starting from the weighted feature matrix, a new remote sensing image sequence is regenerated using steps such as temporal slicing, pooling, generative adversarial network generation, spatial resolution verification, and pixel reconstruction, and breakpoint detection is performed again. This process is repeated, increasing the weight of the spectral fidelity loss function by 10% each time, until the number of breakpoints in the generated new remote sensing image sequence does not exceed a threshold. This sequence is then considered the optimized sequence. This iterative optimization process enhances spectral constraints, gradually improving the continuity and realism of the image sequence, ensuring that the output sequence meets the temporal smoothness requirements of dynamic monitoring.

[0104] In step S18, geometric registration verification is performed based on the optimized sequence to generate a registration deviation matrix. Then, based on the registration deviation matrix and error judgment, iterative correction is performed to obtain the final remote sensing image, including:

[0105] The radiance values ​​and spectral band features of the optimized sequence are extracted to form a set of correction features;

[0106] Based on the set of correction features, spatial transformation parameters are calculated using a pre-constructed geometric correction model to obtain spatial transformation relationships.

[0107] Based on the optimized sequence, the projected coordinates are calculated using the spatial transformation relationship, and a point-by-point difference operation is performed based on the calculation results and the reference coordinates of the spectral band features to obtain the registration deviation matrix.

[0108] When the global root mean square error of the registration deviation matrix is ​​lower than a preset error threshold, the optimized sequence is directly used as the final remote sensing image.

[0109] When the global root mean square error of the registration deviation matrix is ​​not lower than the error threshold, the parameters of the geometric correction model are adjusted, and the spatial transformation relationship is regenerated until the global root mean square error of the registration deviation matrix is ​​lower than the error threshold, thus obtaining the final remote sensing image.

[0110] Specifically, the radiance value and spectral band features of each pixel are extracted from the optimized sequence generated in step S17, and the corresponding spectral band features are extracted from the original remote sensing image obtained in step S11 as a benchmark for geometric correction. The radiance values ​​of the optimized sequence are arranged according to pixel position to form a two-dimensional radiance matrix, and the spectral band features are arranged according to band order to form a multi-dimensional spectral matrix. The radiance value and spectral band features of each pixel are combined into a feature vector, and the feature vectors of all pixels are arranged according to spatial position to form a correction feature set. This set is a three-dimensional data structure with spatial height, spatial width, and feature dimension. The feature dimension is composed of radiance values ​​and reflectance values ​​from multiple spectral bands.

[0111] The set of correction features is input into a pre-constructed geometric correction model for spatial transformation parameter calculation. The geometric correction model employs a polynomial fitting method. Its construction process involves collecting 10,000 samples containing ground control point coordinates from a historical high-precision remote sensing image dataset. Each control point has pixel coordinates on the image to be corrected and true coordinates in the geographic coordinate system. The collected control point coordinate data undergoes preprocessing: points with significant deviations are removed (e.g., by calculating the local density of each control point and removing outliers exceeding a preset threshold); all coordinate values ​​are subtracted from the mean of their corresponding coordinate components and then divided by the standard deviation to ensure that coordinate values ​​of different dimensions are within the same numerical range, improving numerical stability. A second-order polynomial model is used to describe the mapping relationship from pixel coordinates to geographic coordinates, and the model contains six coefficients to be solved. For each sample set, the pixel coordinates and true coordinates of all control points are substituted into the polynomial to construct an overdetermined system of equations. The coefficients are solved using the least squares method, minimizing the sum of squared residuals between the true coordinates of all control points and the polynomial-calculated coordinates. In the specific calculation, the pixel coordinates of the control points are combined into a design matrix, and the real coordinates are combined into an observation vector. The normal equation is then solved to obtain the coefficient vector. After training, a fixed polynomial coefficient is obtained. When using it, the spatial coordinates of each pixel in the calibration feature set are taken as input, substituted into the second-order polynomial, and the geographic projection coordinates corresponding to that pixel are calculated. This process is repeated for all pixels to obtain the projection coordinates of each pixel, forming a projection coordinate matrix.

[0112] The projected coordinates of each pixel in the optimized sequence are differentially analyzed point-by-point with the reference coordinates extracted from the original spectral band features. The reference coordinates are derived from the geolocation information of the original remote sensing image, and each pixel has precise geographic coordinates. For each pixel, the longitude value of its projected coordinates is subtracted from the reference longitude value to obtain the longitude deviation; the latitude value is subtracted from the reference latitude value to obtain the latitude deviation. The longitude and latitude deviations of all pixels are arranged according to their spatial location to form two two-dimensional deviation matrices with the same spatial size as the optimized sequence, namely the longitude deviation matrix and the latitude deviation matrix.

[0113] The global root mean square error (RMSE) is calculated for the longitude and latitude deviation matrices. Specifically, for each pixel, the squares of the longitude and latitude deviations are added together to obtain the squared comprehensive deviation for that pixel. The squared comprehensive deviations of all pixels are then summed and divided by the total number of pixels to obtain the average squared comprehensive deviation. Taking the square root of this average value yields the global root mean square error, which represents the average deviation of the entire image sequence from the reference geographic coordinates.

[0114] The error threshold is determined as follows: 500 images are randomly selected from a historical high-precision registered remote sensing image dataset. Each image has been precisely registered manually, and its global root mean square error (RMSE) is known. The RMSEs of these 500 images are sorted from smallest to largest, and the value at the 95th position is taken as the error threshold. This threshold indicates that, under acceptable registration accuracy, the RMSE of 95% of the high-quality images does not exceed this value.

[0115] The calculated global root mean square error (RMSE) is compared with a preset error threshold. If the RMSE is lower than the threshold, the current optimized sequence is directly output as the final remote sensing image. If the RMSE is not lower than the threshold, the parameters of the geometric correction model need to be adjusted. The adjustment method is to use the deviation between the projected coordinates and the reference coordinates of all pixels in the current correction feature set as feedback to reselect control points. Specifically, one hundred pixels are randomly selected from the region with large deviations as new control points. The pixel coordinates of these control points and their reference coordinates are known. The polynomial coefficients are solved again using the least squares method with these control points to obtain the updated spatial transformation relationship. Then, the projection coordinate calculation and difference operation are re-executed to obtain a new registration deviation matrix and a new RMSE. The above process is repeated, each time reselecting control points and updating polynomial coefficients based on the region with large deviations, until the RMSE is lower than the threshold. When the condition is met, the optimized sequence used in the last iteration is output as the final remote sensing image.

[0116] Reference Figure 2 The second embodiment of the present invention provides a spatiotemporally controllable remote sensing image simulation generation system based on geographic metadata constraints, comprising:

[0117] The data acquisition module is used to acquire topographic elevation values, pixel spatial coordinates, and remote sensing images containing radiance values, spectral band characteristics, and acquisition dates.

[0118] The mapping module is used to perform feature encoding based on the remote sensing image, the terrain elevation value and the pixel spatial coordinates to obtain an abstract feature map, and to perform feature mapping based on the abstract feature map to obtain a geographic information vector.

[0119] The encoding module is used to perform disaster association retrieval through a preset disaster text database based on the geographic information vector to obtain a disaster text report, and to perform semantic encoding based on the disaster text report to obtain regional feature vectors and trajectory feature vectors;

[0120] The alignment module is used to perform dimensional adjustment based on the region feature vector and the trajectory feature vector to obtain the projected spectral vector, and to perform feature space alignment calculation based on the projected spectral vector, the terrain elevation value and the trajectory feature vector to obtain the fused vector group;

[0121] The enhancement module is used to calculate the association weights based on the fusion vector group, obtain the association weight distribution, and perform feature enhancement based on the association weight distribution and weight allocation judgment to obtain a weighted feature matrix.

[0122] The reconstruction module is used to generate images based on the weighted feature matrix using a pre-built generative adversarial network model to obtain image feature maps, and to reconstruct pixels based on the image feature maps and spatial resolution consistency to obtain a remote sensing image sequence.

[0123] The optimization module is used to perform interruption point detection and statistics based on the remote sensing image sequence, obtain the number of interruption points, and perform iterative optimization on the remote sensing image sequence based on the number of interruption points to obtain an optimized sequence.

[0124] The output module is used to perform geometric registration verification based on the optimized sequence, generate a registration deviation matrix, and perform iterative correction based on the error judgment according to the registration deviation matrix to obtain the final remote sensing image.

[0125] It should be noted that the spatiotemporally controllable remote sensing image simulation generation system based on geographic metadata constraints provided in this embodiment of the invention is used to execute all the process steps of the spatiotemporally controllable remote sensing image simulation generation method based on geographic metadata constraints in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0126] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0127] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for generating a spatio-temporally controllable remote sensing image simulation based on geographical meta-information constraints, characterized in that, include: Acquire topographic elevation values, pixel spatial coordinates, and remote sensing images containing radiance values, spectral band characteristics, and acquisition dates; Based on the remote sensing image, the terrain elevation value, and the pixel spatial coordinates, feature encoding is performed to obtain an abstract feature map, and feature mapping is performed based on the abstract feature map to obtain a geographic information vector. Based on the geographic information vector, disaster association retrieval is performed through a preset disaster text database to obtain a disaster text report. Semantic encoding is then performed on the disaster text report to obtain regional feature vectors and trajectory feature vectors. Based on the region feature vector and the trajectory feature vector, dimensional adjustment is performed to obtain the projected spectral vector. Then, based on the projected spectral vector, the terrain elevation value, and the trajectory feature vector, feature space alignment calculation is performed to obtain the fused vector group. Based on the fusion vector group, the association weight is calculated to obtain the association weight distribution. Based on the association weight distribution, feature enhancement is performed based on the weight allocation judgment to obtain the weighted feature matrix. Based on the weighted feature matrix, an image is generated using a pre-constructed generative adversarial network model to obtain an image feature map. Then, based on the image feature map, pixel reconstruction is performed based on spatial resolution consistency to obtain a remote sensing image sequence. Based on the remote sensing image sequence, interruption point detection and statistics are performed to obtain the number of interruption points. Based on the number of interruption points, the remote sensing image sequence is iteratively optimized to obtain an optimized sequence. Based on the optimized sequence, geometric registration verification is performed to generate a registration deviation matrix. Then, based on the registration deviation matrix and error judgment, iterative correction is performed to obtain the final remote sensing image. 2.The method of claim 1, wherein, The step involves performing feature encoding based on the remote sensing image, the terrain elevation values, and the pixel spatial coordinates to obtain an abstract feature map, and then performing feature mapping based on the abstract feature map to obtain a geographic information vector, including: Based on the radiance value and spectral band features in the remote sensing image, as well as the terrain elevation value and the pixel spatial coordinates, feature stitching is performed to generate a hybrid feature tensor. Based on the hybrid feature tensor, feature encoding is performed using a pre-built convolutional neural network model to obtain an abstract feature map; Based on the abstract feature map, feature mapping is performed using pre-constructed map mapping rules to obtain high-dimensional feature vectors; Based on the acquisition date in the remote sensing image, a time series index is obtained by numerical conversion using a pre-constructed time axis encoding rule; Based on the high-dimensional feature vector and the time series index, feature concatenation is performed to generate a geographic information vector containing the time series index, land cover type, and spectral feature vector. 3.The method of claim 2, wherein, The step involves performing disaster association retrieval using a pre-defined disaster text database based on the geographic information vector to obtain a disaster text report, and then performing semantic encoding on the disaster text report to obtain regional feature vectors and trajectory feature vectors, including: Based on the time series index and land cover type in the geographic information vector, a disaster text report is obtained by performing a correlation retrieval through a preset disaster text database; Identify the geographical location nouns and state change verbs in the disaster text report, and extract the descriptive sentences based on the geographical location nouns and state change verbs through dependency parsing. Based on the description, semantic encoding is performed using a pre-built BERT model to obtain region feature vectors and trajectory feature vectors. 4.The method of claim 2, wherein, The process involves adjusting the dimensions based on the region feature vector and the trajectory feature vector to obtain a projected spectral vector, and then performing feature space alignment calculations based on the projected spectral vector, the terrain elevation value, and the trajectory feature vector to obtain a fused vector set, including: When the dimensions of the spectral feature vector and the regional feature vector of the geographic information vector are inconsistent, the dimensional transformation of the spectral feature vector is performed through a pre-constructed linear projection layer to obtain the projected spectral vector; When the spectral feature vector of the geographic information vector and the dimension of the regional feature vector are the same, the spectral feature vector is directly used as the projected spectral vector. Based on the terrain elevation values, a terrain vector is obtained by performing dimensional mapping through a pre-constructed terrain mapping layer, and a trajectory vector is obtained by performing dimensional mapping through the linear projection layer based on the trajectory feature vector. The alignment point is obtained by calculating the average vector of the projected spectral vector, the terrain vector, the regional feature vector, and the trajectory vector; The differences between the projected spectral vector, the terrain vector, the regional feature vector, and the trajectory vector and the alignment point are calculated respectively, and the vectors are concatenated according to the calculation results to obtain a fused vector group containing texture features and text features. 5.The method of claim 4, wherein, The step involves calculating association weights based on the fused vector group to obtain an association weight distribution, and then performing feature enhancement based on the weight allocation judgment according to the association weight distribution to obtain a weighted feature matrix, including: Based on the text features in the fused vector group, an attention query vector is obtained by linear transformation through a pre-constructed query projection layer. Based on the texture features in the fused vector group, linear transformations are performed through a pre-constructed key projection layer and value projection layer to obtain the key matrix and value matrix, respectively. Based on the attention query vector and the key matrix, the weights are calculated using a scaled dot product attention mechanism to obtain the associated weight distribution; Based on the associated weight distribution, the texture features and weight values ​​corresponding to the dimensions whose weight values ​​exceed a preset weight threshold are marked to obtain highly associated feature pairs; Based on the highly correlated features, an amplification factor is calculated using a preset exponential mapping function to obtain a feature enhancement factor. The feature enhancement factor and the value matrix are then multiplied by weight to obtain a weighted feature matrix.

6. The spatiotemporally controllable remote sensing image simulation generation method based on geographic metadata constraints according to claim 2, characterized in that, The process of generating an image using a pre-constructed generative adversarial network model based on the weighted feature matrix to obtain an image feature map, and then reconstructing pixels based on the image feature map and spatial resolution consistency to obtain a remote sensing image sequence, includes: Based on the time series index, the weighted feature matrix is ​​sliced ​​in time series to obtain a time phase slice group. Based on the time phase slice group, spatial features are aggregated by global average pooling to obtain a time phase feature vector group. The difference and cosine similarity between adjacent time-phase feature vectors in the time-phase feature vector group are calculated, and image features are generated through a pre-constructed generative adversarial network model based on the calculation results to obtain an image feature map. When the image feature map satisfies the preset spatial resolution consistency constraint, the spatial pixel dimension is transformed through the pre-constructed pixel reconstruction mapping layer to obtain the pixel reconstruction matrix; When the image feature map does not satisfy the spatial resolution consistency constraint, the parameters of the generative adversarial network model are adjusted, and the image feature map is regenerated until the image feature map satisfies the spatial resolution consistency constraint, thus obtaining the pixel reconstruction matrix; Based on the pixel reconstruction matrix and the image feature map, pixel transformation is performed through matrix multiplication to obtain a remote sensing image sequence.

7. The spatiotemporally controllable remote sensing image simulation generation method based on geographic metadata constraints according to claim 1, characterized in that, The step involves detecting and counting breakpoints in the remote sensing image sequence to obtain the number of breakpoints, and then iteratively optimizing the remote sensing image sequence based on the number of breakpoints to obtain an optimized sequence, including: Based on the remote sensing image sequence and the preset reference surface reflectance, the difference between the pixel value and the reference value is calculated pixel by pixel to obtain the error matrix; Based on the error matrix, breakpoints are detected using a pre-built smoothness detection model to obtain the number of breakpoints. When the number of interruption points is not higher than the preset threshold for the number of interruption points, the remote sensing image sequence is directly used as the optimized sequence. When the number of interruption points exceeds the interruption point threshold, the weights of the spectral fidelity loss function in the generative adversarial network model are adjusted, and the image feature map generation step is re-executed until the number of interruption points is no higher than the interruption point threshold, thus obtaining an optimized sequence.

8. The spatiotemporally controllable remote sensing image simulation generation method based on geographic metadata constraints according to claim 1, characterized in that, The process of performing geometric registration verification based on the optimized sequence, generating a registration deviation matrix, and performing iterative correction based on the registration deviation matrix and error judgment to obtain the final remote sensing image includes: The radiance values ​​and spectral band features of the optimized sequence are extracted to form a set of correction features; Based on the set of correction features, spatial transformation parameters are calculated using a pre-constructed geometric correction model to obtain spatial transformation relationships. Based on the optimized sequence, the projected coordinates are calculated using the spatial transformation relationship, and a point-by-point difference operation is performed based on the calculation results and the reference coordinates of the spectral band features to obtain the registration deviation matrix. When the global root mean square error of the registration deviation matrix is ​​lower than a preset error threshold, the optimized sequence is directly used as the final remote sensing image. When the global root mean square error of the registration deviation matrix is ​​not lower than the error threshold, the parameters of the geometric correction model are adjusted, and the spatial transformation relationship is regenerated until the global root mean square error of the registration deviation matrix is ​​lower than the error threshold, thus obtaining the final remote sensing image.

9. A spatiotemporally controllable remote sensing image simulation generation system based on geographic metadata constraints, characterized in that, include: The data acquisition module is used to acquire topographic elevation values, pixel spatial coordinates, and remote sensing images containing radiance values, spectral band characteristics, and acquisition dates. The mapping module is used to perform feature encoding based on the remote sensing image, the terrain elevation value and the pixel spatial coordinates to obtain an abstract feature map, and to perform feature mapping based on the abstract feature map to obtain a geographic information vector. The encoding module is used to perform disaster association retrieval through a preset disaster text database based on the geographic information vector to obtain a disaster text report, and to perform semantic encoding based on the disaster text report to obtain regional feature vectors and trajectory feature vectors; The alignment module is used to perform dimensional adjustment based on the region feature vector and the trajectory feature vector to obtain the projected spectral vector, and to perform feature space alignment calculation based on the projected spectral vector, the terrain elevation value and the trajectory feature vector to obtain the fused vector group; The enhancement module is used to calculate the association weights based on the fusion vector group, obtain the association weight distribution, and perform feature enhancement based on the association weight distribution and weight allocation judgment to obtain a weighted feature matrix. The reconstruction module is used to generate images based on the weighted feature matrix using a pre-built generative adversarial network model to obtain image feature maps, and to reconstruct pixels based on the image feature maps and spatial resolution consistency to obtain a remote sensing image sequence. The optimization module is used to perform interruption point detection and statistics based on the remote sensing image sequence, obtain the number of interruption points, and perform iterative optimization on the remote sensing image sequence based on the number of interruption points to obtain an optimized sequence. The output module is used to perform geometric registration verification based on the optimized sequence, generate a registration deviation matrix, and perform iterative correction based on the error judgment according to the registration deviation matrix to obtain the final remote sensing image.

Images